@voybio/ace-swarm 2.4.1 → 2.4.2

package/dist/tui/chat.js

@@ -13,6 +13,7 @@ import { ModelBridge } from "../model-bridge.js";
 import { resolveAceStateLayout } from "../ace-state-resolver.js";
 import { applyEvidenceGuardrail, buildAcePreflightPacket, buildBridgeTaskInput, buildContinuityRecord, buildStartupNudge, mapBridgeResultToRuntimeStatus, nextActivationLedger, } from "./local-model-contract.js";
 import { withLocalModelRuntimeRepository, } from "../store/repositories/local-model-runtime-repository.js";
+import { parseExecutionEngine, runLocalModelTask } from "../local-model-runtime.js";
 export class ChatSession extends EventEmitter {
     clients;
     telemetry;
@@ -32,7 +33,10 @@ export class ChatSession extends EventEmitter {
     aceRole;
     aceTier;
     maxTurns;
+    executionEngine;
     aceBridge;
+    hermesExecutor;
+    hermesLaunchProfile;
     activeAceBridge = false;
     providerBaseUrls;
     sessionId = randomUUID();
@@ -56,7 +60,10 @@ export class ChatSession extends EventEmitter {
         this.aceRole = options.aceRole?.trim() || "orchestrator";
         this.aceTier = options.aceTier;
         this.maxTurns = options.maxTurns ?? 6;
+        this.executionEngine = parseExecutionEngine(options.engine) ?? "direct";
         this.aceBridge = options.bridge ?? new ModelBridge(clients);
+        this.hermesExecutor = options.hermesExecutor;
+        this.hermesLaunchProfile = options.hermesLaunchProfile;
         // Add system prompt if provided
         if (this.systemPrompt) {
             this.messages.push({ role: "system", content: this.systemPrompt });
@@ -349,6 +356,9 @@ export class ChatSession extends EventEmitter {
             recommended_next_action: input.recommendedNextAction,
             blocked_reason: input.blockedReason,
             updated_at: Date.now(),
+            execution_engine: this.executionEngine,
+            underlying_provider: this.provider,
+            underlying_model: this.model,
         };
     }
     setRuntimeStatus(status) {
@@ -536,6 +546,70 @@ export class ChatSession extends EventEmitter {
             this.activationLedger = activationLedger;
             return;
         }
+        if (this.executionEngine === "hermes_local") {
+            const delegated = await runLocalModelTask({
+                task: buildBridgeTaskInput(recentConversation, preflight),
+                role: executionRole,
+                workspaceRoot: this.workspaceRoot,
+                provider: streamProvider,
+                model: streamModel,
+                baseUrl: this.providerBaseUrls[streamProvider],
+                ollamaUrl: this.providerBaseUrls.ollama,
+                engine: "hermes_local",
+                maxTurns: this.maxTurns,
+                tier: this.aceTier ?? (executionRole === "orchestrator" ? "compressed" : "brief"),
+                hermesExecutor: this.hermesExecutor,
+                hermesLaunchProfile: this.hermesLaunchProfile,
+            });
+            const assistantText = delegated.result.summary.trim() || "Hermes-local turn completed.";
+            this.messages.push({ role: "assistant", content: assistantText });
+            this.displayMessages.push({
+                role: "assistant",
+                content: assistantText,
+                timestamp: Date.now(),
+                tokens: estimateTokenCount(assistantText),
+            });
+            runtimeStatus = {
+                ...runtimeStatus,
+                bridge_status: delegated.result.status === "completed"
+                    ? "done"
+                    : delegated.result.status === "failed"
+                        ? "failed"
+                        : delegated.result.status === "max_turns"
+                            ? "needs_input"
+                            : delegated.result.status,
+                blocked_reason: delegated.result.status === "failed" ? assistantText : undefined,
+                hermes_bridge_protocol_version: delegated.hermes?.bridge_protocol_version,
+                hermes_session_id: delegated.hermes?.hermes_session_id,
+                shadow_mcp_session_id: delegated.hermes?.shadow_mcp_session_id,
+                updated_at: Date.now(),
+            };
+            this.setRuntimeStatus(runtimeStatus);
+            const continuity = {
+                ...buildContinuityRecord({
+                    preflight,
+                    role: executionRole,
+                    bridgeStatus: runtimeStatus.bridge_status,
+                    evidenceRefs: delegated.result.evidence_refs ?? [],
+                }),
+                execution_engine: "hermes_local",
+                underlying_provider: streamProvider,
+                underlying_model: streamModel,
+                hermes_bridge_protocol_version: delegated.hermes?.bridge_protocol_version,
+                hermes_session_id: delegated.hermes?.hermes_session_id,
+                shadow_mcp_session_id: delegated.hermes?.shadow_mcp_session_id,
+            };
+            this.persistRuntimeSessionArtifacts({ activationLedger, runtimeStatus, continuity });
+            this.activationLedger = activationLedger;
+            this.telemetry.recordRequest({
+                startTime,
+                endTime: Date.now(),
+                promptTokens: estimateTokenCount(text),
+                completionTokens: estimateTokenCount(assistantText),
+                model: streamModel,
+            });
+            return;
+        }
         let bridgeOutput = "";
         const toolNames = [];
         let retryAttempt = 0;

package/dist/tui/index.d.ts

@@ -5,6 +5,7 @@
  * This is the process entry when `ace tui` is invoked.
  */
 import { type TuiController } from "./commands.js";
+import type { HermesLaunchProfile } from "../hermes/launch-profile.js";
 export declare class AceTui implements TuiController {
     private input;
     private layout;
@@ -18,6 +19,7 @@ export declare class AceTui implements TuiController {
     private chatSessions;
     private activeChatSession;
     private provider;
+    private executionEngine;
     private model;
     private ollamaAvailable;
     private providers;
@@ -32,6 +34,7 @@ export declare class AceTui implements TuiController {
     private providerBaseUrls;
     private latestRuntimeStatus;
     private pendingCloseTabId;
+    private hermesLaunchProfile?;
     constructor(options?: {
         provider?: string;
         model?: string;
@@ -40,6 +43,8 @@ export declare class AceTui implements TuiController {
         baseUrl?: string;
         ollamaUrl?: string;
         providerBaseUrls?: Record<string, string>;
+        engine?: string;
+        hermesLaunchProfile?: HermesLaunchProfile;
         workspaceRoot?: string;
         systemPrompt?: string;
     });
@@ -114,6 +119,8 @@ export declare function runTui(options?: {
     baseUrl?: string;
     ollamaUrl?: string;
     providerBaseUrls?: Record<string, string>;
+    engine?: string;
+    hermesLaunchProfile?: HermesLaunchProfile;
     workspaceRoot?: string;
 }): Promise<void>;
 //# sourceMappingURL=index.d.ts.map

package/dist/tui/index.js

@@ -16,8 +16,9 @@ import { ChatSession } from "./chat.js";
 import { OpenAICompatibleClient, diagnoseChatRuntimeConfig, } from "./openai-compatible.js";
 import { detectColorLevel, write, cursor, screen, fg, style } from "./renderer.js";
 import { ALL_AGENTS, WORKSPACE_ROOT } from "../helpers.js";
+import { parseExecutionEngine } from "../local-model-runtime.js";
 import { backfillHandoffsIntoScheduler } from "../tools-handoff.js";
-import { DEFAULT_OLLAMA_MODEL, defaultModelForProvider, inferProviderFromModel, normalizeLocalBaseUrl, } from "./provider-discovery.js";
+import { DEFAULT_OLLAMA_MODEL, defaultModelForProvider, inferProviderFromModel, normalizeLocalBaseUrl, scanLocalModelRuntimes, } from "./provider-discovery.js";
 import { resolveAceStateLayout } from "../ace-state-resolver.js";
 import { withLocalModelRuntimeRepository, } from "../store/repositories/local-model-runtime-repository.js";
 const DASHBOARD_CONTROLS = ["provider", "model", "chat", "logs", "refresh"];
@@ -38,6 +39,7 @@ export class AceTui {
     activeChatSession = null;
     // State
     provider;
+    executionEngine;
     model;
     ollamaAvailable = false;
     providers = [];
@@ -52,10 +54,13 @@ export class AceTui {
     providerBaseUrls = new Map();
     latestRuntimeStatus = null;
     pendingCloseTabId = null;
+    hermesLaunchProfile;
     constructor(options = {}) {
         const workspaceRoot = options.workspaceRoot ?? WORKSPACE_ROOT;
         this.workspaceRoot = workspaceRoot;
+        this.hermesLaunchProfile = options.hermesLaunchProfile;
         this.provider = this.normalizeProvider(options.provider ?? inferProviderFromModel(options.model) ?? "ollama") ?? "ollama";
+        this.executionEngine = parseExecutionEngine(options.engine) ?? "direct";
         const initialModel = options.model ??
             (this.provider === "ollama"
                 ? defaultModelForProvider(this.provider)
@@ -92,6 +97,7 @@ export class AceTui {
             this.config.set("ollama_url", this.providerBaseUrls.get("ollama"));
         }
         this.config.set("provider", this.provider);
+        this.config.set("engine", this.executionEngine);
         this.config.set("model", this.model);
         this.config.set("temperature", "0.7");
         this.config.set("top_p", "0.9");
@@ -148,6 +154,31 @@ export class AceTui {
             this.showMessage(`Provider '${this.provider}' loaded from session/config. Switch with /provider or dashboard controls.`, "info");
             this.checkProviderRuntimeConfig(this.provider, this.model, "startup");
         }
+        // Additionally probe local runtimes (llama.cpp / Ollama) and merge discoveries into TUI so users can explicitly pick either.
+        try {
+            const scan = await scanLocalModelRuntimes({ workspaceRoot: this.workspaceRoot, timeoutMs: 800 });
+            for (const c of scan.candidates) {
+                this.ensureProvider(c.provider);
+                this.setProviderBaseUrl(c.provider, c.baseUrl);
+                if (c.models && c.models.length > 0)
+                    this.setProviderModels(c.provider, c.models, false);
+            }
+            // If the current model appears served by a discovered runtime and the provider wasn't explicitly forced, prefer the local runtime.
+            if (!process.env.ACE_TUI_PROVIDER) {
+                for (const c of scan.candidates) {
+                    if (c.models.includes(this.model) && this.provider !== c.provider) {
+                        this.provider = c.provider;
+                        this.config.set("provider", this.provider);
+                        this.telemetry.setModel(this.model);
+                        this.showMessage(`Discovered local runtime '${c.provider}' serving model '${this.model}'. Provider set to '${this.provider}'.`, "info");
+                        break;
+                    }
+                }
+            }
+        }
+        catch {
+            // Ignore scan failures — discovery is best-effort.
+        }
         // Initial render
         this.fullRender();
         // Start 1-second tick for status bar updates
@@ -498,6 +529,7 @@ export class AceTui {
         const state = {
             taskState: this.getTaskState(),
             provider: this.provider,
+            engine: this.executionEngine,
             model: this.model,
             tokensIn: snap.tokensIn + agentTokens.input,
             tokensOut: snap.tokensOut + agentTokens.output,
@@ -928,6 +960,7 @@ export class AceTui {
         // Create a chat session for this tab
         const session = new ChatSession({ ollama: this.ollama, openai: this.openai }, this.telemetry, {
             provider: this.provider,
+            engine: this.executionEngine,
             model: this.model,
             providerBaseUrls: Object.fromEntries(this.providerBaseUrls.entries()),
             systemPrompt: this.config.get("system_prompt"),
@@ -935,6 +968,7 @@ export class AceTui {
             topP: parseFloat(this.config.get("top_p") ?? "0.9"),
             numCtx: parseInt(this.config.get("num_ctx") ?? "8192", 10),
             workspaceRoot: this.workspaceRoot,
+            hermesLaunchProfile: this.hermesLaunchProfile,
         });
         // Wire chat session events
         session.on("updated", () => {

package/dist/tui/layout.d.ts

@@ -20,6 +20,7 @@ export interface LayoutZones {
 export interface StatusBarState {
     taskState: "idle" | "running" | "blocked" | "overdue";
     provider: string;
+    engine?: string;
     model: string;
     tokensIn: number;
     tokensOut: number;

package/dist/tui/layout.js

@@ -160,6 +160,9 @@ export class LayoutManager {
             ? `${fg.gray}tab: ${fg.white}${state.activeTabLabel}${style.reset}`
             : undefined;
         const providerSeg = `${fg.cyan}provider: ${fg.brightCyan}${state.provider}${style.reset}`;
+        const engineSeg = state.engine
+            ? `${fg.cyan}engine: ${fg.brightCyan}${state.engine}${style.reset}`
+            : undefined;
         const modelSeg = `${fg.cyan}model: ${fg.brightCyan}${state.model}${style.reset}`;
         const runtimeSummary = formatRuntimeStatusSummary(state);
         const runtimeSeg = runtimeSummary
@@ -175,7 +178,7 @@ export class LayoutManager {
                 : state.bridgeStatus
                     ? `${fg.gray}${state.bridgeStatus}${style.reset}`
                     : undefined;
-        const segments = [stateIndicator, tabSeg, providerSeg, modelSeg, runtimeSeg, contextSeg, tokenSeg, timeSeg, clockSeg]
+        const segments = [stateIndicator, tabSeg, providerSeg, engineSeg, modelSeg, runtimeSeg, contextSeg, tokenSeg, timeSeg, clockSeg]
             .filter((segment) => Boolean(segment));
         const sep = ` ${fg.gray}│${style.reset} `;
         const line = ` ${segments.join(sep)} `;

package/dist/tui/provider-discovery.js

@@ -72,22 +72,19 @@ export function inferProviderFromModel(model) {
     const value = model.trim().toLowerCase();
     if (!value)
         return undefined;
-    if (value.startsWith("llama.cpp/") || value.startsWith("llamacpp/")) {
+    // explicit prefixes
+    if (value.startsWith("llama.cpp/") || value.startsWith("llamacpp/"))
         return "llama.cpp";
-    }
-    if (value.startsWith("ollama/") ||
-        value.includes("llama") ||
-        value.includes("qwen") ||
-        value.includes("mistral") ||
-        value.includes("deepseek") ||
-        value.includes("phi") ||
-        value.includes(":") ||
-        value.includes("mixtral")) {
+    if (value.startsWith("ollama/"))
         return "ollama";
-    }
-    if (value.startsWith("copilot/")) {
+    if (value.startsWith("copilot/"))
         return "copilot";
-    }
+    // file-like or gguf artifacts -> prefer llama.cpp
+    if (value.endsWith(".gguf") || value.includes(".gguf"))
+        return "llama.cpp";
+    if (value.endsWith(".bin"))
+        return "llama.cpp";
+    // common hosted provider hints
     if (value.includes("claude"))
         return "claude";
     if (value.includes("gemini"))
@@ -100,6 +97,11 @@ export function inferProviderFromModel(model) {
         value.startsWith("o5")) {
         return "codex";
     }
+    // model families often served via Ollama
+    if (value.includes("qwen") || value.includes("mistral") || value.includes("mixtral") || value.includes("deepseek") || value.includes("phi")) {
+        return "ollama";
+    }
+    // Ambiguous cases (e.g., hf.co/owner/name:ref) should not default to Ollama; allow runtime discovery to decide.
     return undefined;
 }
 export function isLocalLlmProvider(providerInput) {

package/dist/vericify-bridge.d.ts

@@ -14,6 +14,7 @@ export interface VericifyProcessPost {
     agent_id: string;
     kind: VericifyProcessPostKind;
     summary: string;
+    blocker_category?: string;
     tool_refs: string[];
     evidence_refs: string[];
     checkpoint_ref?: string;
@@ -109,6 +110,7 @@ export declare function appendVericifyProcessPost(input: {
     agent_id: string;
     kind: VericifyProcessPostKind;
     summary: string;
+    blocker_category?: string;
     tool_refs?: string[];
     evidence_refs?: string[];
     checkpoint_ref?: string;
@@ -124,6 +126,7 @@ export declare function appendVericifyProcessPostSafe(input: {
     agent_id: string;
     kind: VericifyProcessPostKind;
     summary: string;
+    blocker_category?: string;
     tool_refs?: string[];
     evidence_refs?: string[];
     checkpoint_ref?: string;

package/dist/vericify-bridge.js

@@ -50,6 +50,7 @@ function normalizePost(input) {
         agent_id: input.agent_id.trim(),
         kind: input.kind,
         summary: input.summary.trim(),
+        blocker_category: input.blocker_category?.trim() || undefined,
         tool_refs: [...new Set(input.tool_refs.map((entry) => entry.trim()).filter(Boolean))],
         evidence_refs: [
             ...new Set(input.evidence_refs.map((entry) => entry.trim()).filter(Boolean)),
@@ -354,6 +355,7 @@ export async function appendVericifyProcessPost(input) {
         agent_id: agentId,
         kind: input.kind,
         summary,
+        blocker_category: input.blocker_category,
         tool_refs: input.tool_refs ?? [],
         evidence_refs: input.evidence_refs ?? [],
         checkpoint_ref: input.checkpoint_ref,
@@ -375,6 +377,7 @@ export async function appendVericifyProcessPost(input) {
                         metadata: {
                             branch_id: post.branch_id,
                             lane_id: post.lane_id,
+                            blocker_category: post.blocker_category,
                             checkpoint_ref: post.checkpoint_ref,
                             tool_refs: post.tool_refs,
                             evidence_refs: post.evidence_refs,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@voybio/ace-swarm",
-  "version": "2.4.1",
+  "version": "2.4.2",
   "description": "ACE Framework MCP server and CLI — single-file ACEPACK state, local-model serving, agent orchestration, and host compliance enforcement.",
   "type": "module",
   "main": "dist/index.js",
@@ -21,6 +21,7 @@
   "files": [
     "dist/**/*.js",
     "dist/**/*.d.ts",
+    "scripts/hermes_bridge_worker.py",
     "assets",
     "README.md",
     "CHANGELOG.md"

package/scripts/hermes_bridge_worker.py

@@ -0,0 +1,136 @@
+#!/usr/bin/env python3
+"""ACE-owned Hermes bridge worker.
+
+The worker reads one JSON request from stdin and writes newline-delimited
+bridge events to stdout. All Hermes stdout chatter is redirected to stderr so
+ACE never has to guess whether an unframed line is assistant text or authority.
+"""
+
+from __future__ import annotations
+
+import contextlib
+import json
+import os
+import sys
+import traceback
+from pathlib import Path
+from typing import Any
+
+
+ORIGINAL_STDOUT = sys.stdout
+
+
+def emit(event: dict[str, Any]) -> None:
+    ORIGINAL_STDOUT.write(json.dumps(event, ensure_ascii=False, separators=(",", ":")) + "\n")
+    ORIGINAL_STDOUT.flush()
+
+
+def load_request() -> dict[str, Any]:
+    raw = sys.stdin.read()
+    if not raw.strip():
+        raise ValueError("empty bridge request")
+    parsed = json.loads(raw)
+    if not isinstance(parsed, dict):
+        raise ValueError("bridge request must be a JSON object")
+    return parsed
+
+
+def provider_for_hermes(provider: str) -> str:
+    if provider in {"llama.cpp", "ollama"}:
+        return "openai"
+    return provider
+
+
+def base_url_for_hermes(provider: str, base_url: Any) -> str:
+    raw = str(base_url or "").rstrip("/")
+    if provider == "ollama" and raw and not raw.endswith("/v1"):
+        return f"{raw}/v1"
+    return raw
+
+
+def main() -> int:
+    try:
+        request = load_request()
+        hermes_root = Path(request["hermes_root"]).resolve()
+        sys.path.insert(0, str(hermes_root))
+
+        session_id = str(request["session_id"])
+        turn_id = str(request.get("turn_id") or session_id)
+        emit({"type": "session_open", "session_id": session_id, "turn_id": turn_id})
+
+        with contextlib.redirect_stdout(sys.stderr):
+            from run_agent import AIAgent  # type: ignore
+
+            def status_callback(kind: str, message: str) -> None:
+                emit({"type": "status" if kind == "lifecycle" else "status", "session_id": session_id, "turn_id": turn_id, "kind": kind, "message": str(message)})
+
+            def delta_callback(text: str) -> None:
+                emit({"type": "delta", "session_id": session_id, "turn_id": turn_id, "text": str(text)})
+
+            def reasoning_callback(text: str) -> None:
+                emit({"type": "reasoning", "session_id": session_id, "turn_id": turn_id, "text": str(text)})
+
+            def interim_callback(text: str, **kwargs: Any) -> None:
+                emit({"type": "assistant_interim", "session_id": session_id, "turn_id": turn_id, "text": str(text), **kwargs})
+
+            def thinking_callback(text: str) -> None:
+                emit({"type": "reasoning", "session_id": session_id, "turn_id": turn_id, "text": str(text)})
+
+            def tool_start_callback(tool_name: str, *args: Any, **kwargs: Any) -> None:
+                emit({"type": "tool_start", "session_id": session_id, "turn_id": turn_id, "tool": str(tool_name), "input": kwargs.get("args") or (args[0] if args else {})})
+
+            def tool_progress_callback(tool_name: str, message: str = "", *args: Any, **kwargs: Any) -> None:
+                emit({"type": "tool_progress", "session_id": session_id, "turn_id": turn_id, "tool": str(tool_name), "message": str(message), **kwargs})
+
+            def tool_complete_callback(tool_name: str, result: Any = None, *args: Any, **kwargs: Any) -> None:
+                emit({"type": "tool_complete", "session_id": session_id, "turn_id": turn_id, "tool": str(tool_name), "result": result, **kwargs})
+
+            agent = AIAgent(
+                base_url=base_url_for_hermes(str(request.get("provider") or ""), request.get("base_url")),
+                api_key=str(request.get("api_key") or "no-key-required"),
+                provider=provider_for_hermes(str(request.get("provider") or "")),
+                model=str(request.get("model") or ""),
+                max_iterations=int(request.get("max_turns") or 6),
+                enabled_toolsets=["mcp-ace_shadow"],
+                disabled_toolsets=["terminal", "filesystem", "web", "vision", "creative", "reasoning"],
+                quiet_mode=True,
+                verbose_logging=False,
+                ephemeral_system_prompt=str(request.get("system_prompt") or ""),
+                session_id=session_id,
+                platform="ace",
+                skip_context_files=True,
+                skip_memory=True,
+                status_callback=status_callback,
+                stream_delta_callback=delta_callback,
+                reasoning_callback=reasoning_callback,
+                thinking_callback=thinking_callback,
+                interim_assistant_callback=interim_callback,
+                tool_start_callback=tool_start_callback,
+                tool_progress_callback=tool_progress_callback,
+                tool_complete_callback=tool_complete_callback,
+            )
+            emit({"type": "turn_run", "session_id": session_id, "turn_id": turn_id})
+            result = agent.run_conversation(str(request.get("task") or ""), task_id=turn_id)
+
+        final = result.get("final_response") if isinstance(result, dict) else str(result)
+        emit({
+            "type": "final",
+            "session_id": session_id,
+            "turn_id": turn_id,
+            "text": final or "",
+            "completed": bool(result.get("completed", True)) if isinstance(result, dict) else True,
+            "turns": int(result.get("api_calls", 1) or 1) if isinstance(result, dict) else 1,
+        })
+        emit({"type": "session_close", "session_id": session_id, "turn_id": turn_id})
+        return 0
+    except Exception as exc:
+        emit({
+            "type": "error",
+            "message": str(exc),
+            "traceback": traceback.format_exc(limit=8),
+        })
+        return 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())